Existing device-to-device (D2D) offloading techniques assume that all nodes storing data are potential forwarders. This leads to suboptimal results whenever the system has to reward forwarders. How to design a global strategy that keeps the number of seed users low (to save cellular bandwidth) and selects the appropriate set of forwarders (to know which ones to reward) remains an open issue. We formulate this question as a stochastic control problem that we solve using an application of Pontryagin's Maximum Principle (PMP). We provide a framework that works under a generic cost model. We show analytically that an optimal solution exists and compute when operators benefit from this policy.